Breaking ICD Codes: Identifying Ambiguous Respiratory Infection Codes via Regional Diagnosis Heterogeneity

INTRODUCTION

The World Health Organization (WHO) has emphasized “The goal of having ICD [International Classification of Diseases] in place for health data collection is to generate comparable health data at the international level,”¹ and it intends that this classification will be a “clinically relevant classification system.”¹ The literature pertaining to whether the intended goal has been achieved is relatively sparse in directly addressing this problem, although some studies indirectly provide insight. The issue was theoretically addressed by Cimino² in a highly influential 1998 publication on hierarchical medical terminology. In his later coauthored works, however, limitations in implementing detailed electronic health records were highlighted, including increased physician workload without compelling evidence of any benefit.³ He and colleagues also acknowledged some diagnostic variability, although it arose from differing laboratory thresholds.⁴

There are reports showing some medical variations in the interpretation of symptoms of respiratory infection, although most seem to indicate genuine challenges in diagnosis. Some imply notable controversy among medical specialists regarding diagnoses based on recorded lung sounds, primarily attributed to low specificity and sensitivity in detecting pneumonia or bronchitis based on wheezing and crackles.⁵ Similarly, studies demonstrate a marked enhancement in the accuracy of symptom-based diagnoses when augmented with tests detecting a specific pathogen, as observed in the case of influenza.⁶ A study on international discrepancies of medical coding, however, despite having a relatively small sample, found that respiratory infections were less problematic than other diagnoses.⁷

The anomalies in diagnosis rates could also be seen as a result of a flawed incentive structure. For example, because of isolation requirements under COVID-19, people without paid sick leave were disinclined to obtain a diagnosis.⁸ Reimbursement and supervision mechanisms may also lead to different coding; overly vague codes may potentially lead to claim denial.⁹ Martinez et al¹⁰ noted that the same physicians who are more likely to diagnose potential exceptions such as sinusitis, pharyngitis, or bronchitis are, even after correcting for diagnosis, more likely to prescribe antibiotics for exactly those conditions.

Acute respiratory infection (ARI) represents a cluster of clinically similar conditions for which the International Classification of Diseases, 10th Revision (ICD-10) offers numerous potentially applicable diagnosis codes (codes J00-J22). Clinicians are tasked with selecting among these codes (assigning a diagnosis) based on available clinical, laboratory, and imaging findings for a given patient. In this study, we analyzed comparability of data based on the assignment of ICD-10 codes to ARIs. Our goal was to identify substantial regional differences in the frequency of these codes, indicating heterogeneous diagnoses.

METHODS

RESULTS

The relative prevalences of particular diagnosis codes (diagnoses) applied to ARI for the 380 Polish regions/counties are presented for each age group in Table 1. The data show that a few diagnoses effectively dominate, accompanied by a wide range of their regional proportions. However, this finding alone does not establish whether these differences reflect true epidemiologic variations or merely coding artifacts.

Figure 1 presents correlations between shares of a particular diagnosis, with the existence of strongly negatively correlated codes implying coding ambiguity. In all age groups, the most troublesome code was J06 (acute upper respiratory infections of multiple and unspecified sites). Among children, it was used almost interchangeably with J00 (common cold). For the middle-aged population, J06 was used as an alternative for almost any other upper respiratory infection, and the same was true to a lesser extent among older adults. In each age group, certain infections were designated as either lower respiratory infections (J20 - bronchitis) or upper ones (J00 or J06). Similarly, the same conditions could be diagnosed as either J02 (pharyngitis) or as a more severe condition, especially J10 (influenza due to other identified influenza virus) or J15 (bacterial pneumonia, not elsewhere classified). There was one negative correlation that could plausibly be attributed to differences in use of medical services: a weak negative relationship observed in both adult groups between the common cold and pneumonia.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

Network Plots Showing Correlations of Diagnoses by Age Group

Note: See Table 1 footnotes for code definitions.

Some diagnosis codes showed an interesting subtle positive correlation, particularly noticeable among nuanced lower respiratory infections. Across all groups, J04 (laryngitis and tracheitis) was positively correlated with J11 (influenza due to an unidentified influenza virus), which may seem counterintuitive. As our analysis was conducted before lateral flow tests were issued,¹¹ however, this positive correlation may instead imply the co-occurrence of detailed diagnoses. This mechanism becomes clearer with another pair of diagnoses showing a positive correlation across all groups: J13 (pneumonia due to Streptococcus pneumoniae) and J14 (pneumonia due to Hemophilus influenzae), suggesting that certain facilities implemented more rigorous testing protocols.

In order to ascertain whether the observed relationship was a systemic problem or the wide ranges were merely caused by a minuscule number of miscoding clinicians, we assessed selected cases with the strongest negative correlation for all groups in pairs (Figure 2). Assuming that the same criteria had been consistently applied in each region, the plotted lines would be effectively flat, except for a few outliers in the top and bottom percentile. Nevertheless, although such outliers were detectable, there was no constant segment; this implies that the observed patterns were not due to isolated miscoding but rather reflected a broader lack of consensus on diagnostic delimitation.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

Relative Region (Powiat) Probabilities of Receiving a Particular Diagnosis From Ambiguous Pairs

Note: See Table 1 footnotes for code definitions.

Finally, in order to verify whether any discernible regional pattern could explain the observed differences, we plotted relationships for the pair with the strongest negative correlation, detected for pair J00 and J06 among children (Figure 3). The map did not show any consistent regional differences. The only noticeable relationship was decreasing dispersion with increasing region diagnosis count, suggesting that the observed phenomena are not indicative of better diagnosis in larger cities, but rather that as diagnosis count increases, random factors begin to cancel out.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3.

Relative Region (Powiat) Probabilities of Receiving a Particular Diagnosis for the Pair of J00 and J06 Among Children

J00 = acute nasopharyngitis [common cold]; J06 = acute upper respiratory infections of multiple and unspecified sites; Loess = locally weighted scatterplot smoothing.

DISCUSSION

To the best of our knowledge, this is the first study where observable anomalies were used to analyze the extent of regional variability in coding for the same respiratory infection. The observed differences in ICD-10 diagnosis code assignment for ARI appear huge, and although literature on the subject acknowledges some genuine diagnostic problems, these problems are insufficient to explain the magnitude of mismatch; thus, the most plausible explanation appears to be ambiguity of the codes. The International Classification of Diseases, 11th Revision (ICD-11) is highly unlikely to mitigate this problem as it has yet additional ARI diagnosis codes.¹²